Development of operation and maintenance strategies for offshore wind industry based on big data management

Abstract

Increasingly, enterprises allocate substantial funds to offshore wind energy both development and deployment as a key element of the global energy transition from fossil energies; hence, the importance of ensuring the technical reliability of offshore wind turbines becomes significant for the viability of the industry. Availabilities of offshore wind turbines can be significantly lower, by 15% or more, than the onshore wind turbines, with a typical availability of 95-97%, leading to a reduced electricity production of similar magnitude. In that context, predictive maintenance is an essential tool for increasing offshore wind turbines' operational availability to maximize the energy yield. At the same time, artificial intelligence (AI) is progressively introduced to operation and maintenance (O&M) of offshore wind farms for enhancing the efficiency and performance of the wind energy power plants and projects. An important trend hereby are decision support strategies based on failure predictions. This results in AI being more frequently used to create time-to-failure predictions based on large amount of data collected from sensors deployed to wind turbines. However, components or subsystems that are unsupervised are not covered here and may occasionally lead to failures. The central focus of this work lies in the design, development, and practical application of a prediction method of failures of unsupervised sub-systems in a real-life wind turbine system to increase the overall technical availability of offshore wind turbines utilizing data-driven approaches to validate the proposed methodology. As an example, the brake pads of the yaw system are selected, to prove the methods developed in a real environment. As a result of the research works performed, an AI-based method has been established for predicting component failures which has been verified in practical application taking the yaw brakes of a 3MW wind turbine as an example. It is analyzed how the brake pads of these yaw brakes wear out over time, using the data collected from turbine controllers. To predict when these failures are likely to occur, Long-Short-Term Memory (LSTM) is employed which is empowered by a pre-processed dataset using Support Vector Machine (SVM) for clustering of the relevant data. This combination of SVM and LSTM presents an alternative approach to enhancing predictive maintenance strategies, which can improve the operational reliability and cost-efficiency of offshore wind energy systems. In particular, the presented methodology leverages deep learning models in conjunction with online clustering techniques, thus establishing a robust foundation for predictive maintenance in offshore wind energy systems. Accordingly, a comprehensive framework, methodology, experimental setup, and results are presented here for giving new contributions and suggesting insights into the evolving landscape of predictive maintenance of offshore wind turbines. An overview of the history of offshore wind energy and a classification of offshore wind turbines and wind farms is given including the outlining of challenges related to offshore wind turbine development and operation that fundamentally differs from onshore wind energy in light of design and operations conditions due to for example limited accessibility. Operation and Maintenance is discussed presenting O&M classifications and definitions based on different norms. Decision Support Strategies (DSS) as tool for O&M are introduced and strategies based on Key Performance Indicators developed. The dependence of the different models on input data is presented for maintenance strategies in detail. Conditions that impact operation and maintenance are worked out and clustered specifically for offshore wind energy. A discussion about challenges related to data collection and handling is discussed, which are mainly data availability and the fact that not every component or instance in offshore wind turbines is sufficiently supervised and its data logged. On data utilization for offshore wind energy applications, it is laid out that, for the utilization of data collected in and around offshore wind farms, strategies based on machine learning and artificial intelligence techniques as basis for decision making in operation and maintenance of offshore wind energy must be applied. In a wind farm project several sensors are monitored and logged that can ideally be used for the training and testing procedures of classification algorithms. However, for specific processes, not all of those sensors provide conditional information that are useful to take into consideration, while several of them may be correlated. Therefore, as first step Non- Negative Matrix Factorization and Principal Component Analysis are explained as utilities for performing of feature reduction and dimensionality reduction that need to be performed to narrow down available data to the problem in focus. After this step, data can be classified and scaled to be suitable for data utilization strategies. SVM, as a capable classifier, is explained in detail. Different neural networks, such as artificial neural networks, feed-forward neural networks, and long-short-term memory recurrent neural networks (LSTM), are presented as suitable data utilization strategies, and the steps for data integration into predictive maintenance strategies are explained. A case study is presented, developing a prediction model for an unsupervised component in the yaw system of an existing wind turbine whose degrading leads to problems in the operation of the wind turbine. Specifically, the yaw system of this specific wind turbine model experiences increasing number of unintended slipping events where the yaw system is not able to keep the wind turbine stationary due to unsupervised wear out of brake pads. This leads in consequence to increased yaw activities and reduced service life of the wind turbine. For the case study, real data of over six months have been taken from a 3MW direct-drive wind turbine. The number of variables logged over a period of six months is with a count of 216 at a frequency of 1Hz very high. For this a categorization of the data collected from the turbine is elucidated and feature reduction techniques are applied leading to a reduced number of six features that represent the data to be utilized for predictions. Also, a rigorous evaluation of the selected methodology, namely SVM as classifier in combination with LSTM to forecast the evolution of increasing number of slipping events, as which places significant reliance on the utilization of deep learning models, is provided. This analysis serves to underscore the rationale behind the methodological choices. Comprehensive formulation of the problem and elucidation of underlying assumptions pivotal to predictive maintenance strategies, specifically the inclusion of unsupervised sub-systems and components into predictive maintenance strategies is given in detail. A detailed description of these considerations serves as basis for the subsequent development of prediction strategies. Finally, the results obtained through the application of the selected predictive strategy for unsupervised sub-systems is presented. These results are scrutinized in detail to offer a comprehensive assessment of their implications and significance within the broader context of predictive maintenance for offshore wind turbines. In the evaluation of the results, key findings stemming from the application of the proposed predictive maintenance methodology are laid out. This work and especially the results from the case study show the potential of AI for the further development of predictive maintenance based on large amounts of available data. The case study also proves that not only occurrence of catastrophic failures but also prediction of points in time when error frequency limits are exceeded can be predicted reliably. The clustering capabilities of SVM hereby enhance the performance of LSTM, which is selected to achieve reliable prediction outcomes. In this work a prediction period of 15 days was achieved with an overall prediction accuracy of over 98% making the model sufficient as input for DSS with focus on preventive or reliability centered maintenance strategies. In conclusion this work shows how combining domain expertise with machine learning techniques can improve data based predictive maintenance practices in the offshore wind industry by including previously unsupervised areas that have an impact on turbine availability, reliability, lower operation cost and consequently on power output and feasibility as well as on service life. Applying this method to other unsupervised components or sub-systems for future applications is possible leading to a higher level of confidence on the integrational DSS. The application of a combination of well-developed and in literature comprehensive presented AI methods to the field of offshore wind energy in order to predict unsupervised components, sub-systems or events adds an important novel aspect to the literature.